# I Spent Two Years Building the Median — Hariprasad Sudharshan

## Executive summary

The talk argues that while AI and LLMs have drastically lowered the barrier to entry for software development, making the 'median' solution readily available and free, this convergence means that mere technical capability is no longer a durable product advantage. The speaker proposes that successful products must shift focus from the model's capability (the 'athlete') to the novelty of the product itself, emphasizing the 'novelty budget' and distribution (the 'moat'). Furthermore, he stresses that the underlying global economy still relies on reliable, deterministic software, even in the age of AI.

## Key takeaways

- The World Runs on Deterministic Software: The modern global economy, from infrastructure to finance, is driven by traditional, deterministic, rule-based software, not just AI. AI's primary role is to accelerate the development of this reliable deterministic software at the lowest cost possible.
- The Model, Harness, and Product Distinction: The model is the underlying capability (the engine/athlete). The harness includes tools, context, and orchestration (the training system). The product is the ultimate value proposition—the reason a customer should care. Differentiation must come from the product, not just the model.
- The Median Problem and Novelty Budget: Because LLMs and coding agents make solving common problems (like RAG systems) too easy, the resulting solutions are converging into a 'median.' To compete, developers must dedicate a 'novelty budget'—identifying and building a truly unique element that the model cannot supply.
- Three Questions for Product Leaders: When evaluating a demo, leaders should ask: 1) Who has already solved this? 2) What does our team have that a competitor does not? 3) Where will the application reach users? (Focusing on distribution/moat).

## Technical details

- Software Architecture: The speaker describes the software stack as a cross-feedback loop (e.g., assemblers built C, C built Unix, Unix built Kafka, Zookeeper, Kubernetes). The final artifact, regardless of whether it uses probabilistic AI, must be deterministic and reliable.
- AI Development Workflow: The 'build first reflex' is the tendency to immediately open a coding agent when faced with a problem, making development faster but often leading to building a 'median' solution that is already solved or available in mature open-source systems.
- Model Deployment Costs: A key differentiator is the cost difference between running open model weights inside an enterprise versus purchasing inference via a cloud service (e.g., Neo cloud). This involves understanding the infrastructure layers (e.g., precision KV cache, batching, execution parallelism).

## Practical implications

- When starting a new project, dedicate a 'novelty budget' by identifying one unique, non-obvious element before writing code.
- For leaders, shift focus from evaluating the demo's technical complexity to assessing the distribution channel and competitive advantage.
- Recognize that the most valuable work now is not building the solution, but finding the next great design or the unique problem that the AI agent cannot solve.
- If building for internal consumption, the AI-assisted process is highly valuable; if competing in the open market, novelty is mandatory.

## Topics

AI Software Development, Product Strategy, Software Architecture, LLMs, Build Engineering, GitHub, Claude

Source: https://www.youtube.com/watch?v=0JZy5AuuQzc